MetaSimGrid : Towards realistic scheduling simulation of distributed applications

(eng) Most scheduling problems are already hard on homogeneous platforms, they become quite intractable in an heterogeneous framework such as a metacomputing grid. In the best cases, a guaranteed heuristic can be found, but most of the time, it is not possible. Real experiments or simulations are often involved to test or to compare heuristics. However, on a distributed heterogeneous platform, such experiments are technically difficult to drive, because of the genuine instability of the platform. It is almost impossible to guarantee that a platform which is not dedicated to the experiment, will remain exactly the same between two tests, thereby forbidding any meaningful comparison. Simulations are then used to replace real experiments, so as to ensure the reproducibility of measured data. A key issue is the possibility to run the simulations against a realistic environment. The main idea of trace-based simulation is to record the platform parameters today, and to simulate the algorithms tomorrow, against the recorded data: even though it is not the current load of the platform, it is realistic, because it represents a fair summary of what happened previously. A good example of a trace-based simulation tool is \textscSimGrid, a toolkit providing a set of core abstractions and functionalities that can be used to easily build simulators for specific application domains and/or computing environment topologies. Nevertheless, \textscSimGrid lacks a number of convenient features to craft simulations of a distributed application where scheduling decisions are not taken by a single process. Furthermore, modeling a complex platform by hand is fastidious for a few hosts and is almost impossible for a real grid. This report is a survey on simulation for scheduling evaluation purposes and present \textscMetaSimGrid, a simulator built on top of \textscSimGrid